While performing Adaptive Bias Force method free energy calculation in NAMD we divide our region of PMF (Potential of Mean Force) in different windows and then merge the .pmf files performing another simulation. Why do we do this? How can I know whether my region of PMF should be divided into windows or not? What will be the number of windows I'll need to divide?
Using windows assures more rapid exploration of your transition coordinate and more uniform sampling. Without windows, often you will get millions of samples in one part of your domain and few or none in others. Using windows helps to cover the domain more uniformly (although other techniques, such as multiple replicas, can also help in this regard, see the linked paper). Even once ABF has converged, sampling along the transition coordinate is free diffusion, which means that the distance traveled goes like t^(1/2). So sampling a large domain with a single window can be very slow.
Note that using windows that are too small can also be detrimental because you may constrain the paths that can be taken along coordinates orthogonal to the transition coordinate. For example, consider calculating the potential of mean force of a molecule as a function of distance from an approximately flat surface with two binding pockets. Let's say we have an initial condition that begins in one binding pocket. If the ends of the window don't extend beyond the binding pocket, you will have no chance to sample the other binding pocket. If your window is large enough, the molecule can leave one binding pocket, diffuse about the larger volume above, and perhaps enter the second binding pocket. Of course it would be best to identify this situation and maybe perform multidimensional ABF, but such phenomena are not always so obvious. Having overlapping windows can also be a good strategy, particularly because it allows you to check that two windows individually give similar results in the regions that they coincide.
Article Achieving ergodic sampling using replica-exchange free-energ...
The reason for using the windows which basically a discretization of the reaction coordinate is to transfer your system from one state to another state in a quasistatic fashion. This is advantageous because evaluating the free energy difference between each two consecutive quasistatic states does not require extensive phase space sampling. As for the number of windows used, it is clear that for all systems the lesser the number of windows used the more extensive phase space sampling required to reach equilibrium and accurately evaluate the free energy difference between any two consecutive windows (quasistatic states). Therefore, the number of windows you use will depend on your system, its size, the reaction coordinate used and the simulation time required to achieve equilibrium between any two consecutive windows.
Using windows assures more rapid exploration of your transition coordinate and more uniform sampling. Without windows, often you will get millions of samples in one part of your domain and few or none in others. Using windows helps to cover the domain more uniformly (although other techniques, such as multiple replicas, can also help in this regard, see the linked paper). Even once ABF has converged, sampling along the transition coordinate is free diffusion, which means that the distance traveled goes like t^(1/2). So sampling a large domain with a single window can be very slow.
Note that using windows that are too small can also be detrimental because you may constrain the paths that can be taken along coordinates orthogonal to the transition coordinate. For example, consider calculating the potential of mean force of a molecule as a function of distance from an approximately flat surface with two binding pockets. Let's say we have an initial condition that begins in one binding pocket. If the ends of the window don't extend beyond the binding pocket, you will have no chance to sample the other binding pocket. If your window is large enough, the molecule can leave one binding pocket, diffuse about the larger volume above, and perhaps enter the second binding pocket. Of course it would be best to identify this situation and maybe perform multidimensional ABF, but such phenomena are not always so obvious. Having overlapping windows can also be a good strategy, particularly because it allows you to check that two windows individually give similar results in the regions that they coincide.
Article Achieving ergodic sampling using replica-exchange free-energ...
- Badges
- Science topic
- Similar topics
- Chemistry
- Theoretical Chemistry
- Computational Chemistry
- Molecular Dynamics